METR Benchmark: leaderboard
189 software engineering, ML, and cybersecurity tasks measuring AI agents' ability to autonomously complete long-horizon tasks, scored by task-completion time horizon.
Metric: 50% Time Horizon (hours). Source: metr.org. Saturation forecast: Around August 2027. 26 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Mythos Preview | 17.41 | #4 |
| 2 | Claude Opus 4.6 | 11.98 | #63 |
| 3 | Gemini 3.1 Pro (Preview) | 6.4 | #58 |
| 4 | GPT-5.2 | 5.87 | #110 |
| 5 | GPT-5.3 Codex | 5.83 | #69 |
| 6 | GPT-5.4 | 5.7 | #79 |
| 7 | Claude Opus 4.5 | 4.88 | #83 |
| 8 | Gemini 3 Pro | 3.74 | #80 |
| 9 | GPT-5.1 Codex Max | 3.73 | #100 |
| 10 | GPT-5 | 3.38 | #98 |
| 11 | O3 | 2 | #124 |
| 12 | Claude Opus 4 | 1.67 | #163 |
| 13 | Claude Opus 4.1 | 1.67 | #140 |
| 14 | Claude 3.7 Sonnet | 1.01 | #249 |
| 15 | O1 | 0.65 | #183 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=metr-benchmark · How It Works · Data refreshed daily, snapshot 2026-10-20.